Unexpected behaviour of Stochastics/Deterministic variables in an 'if' statement.

[mar-ses]

def forward_radius(Flux, Mass, A_const=A, C_const=C, sw=log_switchpoint):
    if(Flux < 10**sw.value):
        muR = C_const
        print('THERE')
    else:
        muR = A_const*log10(Flux) + C_const - A_const*sw
        print('HERE')
    return muR

R = np.empty(N, dtype=object)

for i in range(N):
    R[i] = pymc.Normal('R_%i' %i, mu = forward_radius(F[i], M[i], A, C, log_switchpoint), tau=tauR)

Hello, in the above example is an excerpt of my code. I was having a problem so I simplified the code to see exactly what was wrong and I figured it out.
R, F and M are arrays of length N containing pymc.Stochastic objects. A and C are hyperparameters (also pymc.Stochastic objects with a prior distribution).
The forward model is R[i] ~ Norm( mu = f(M[i], F[i], A, C, switchpoint), tau = ... )

The problem is that when the forward_radius() function is entered, the else: clause never seems to be activated. I put the print statements there to check when each is entered and when. Apparently what happens is that the condition only seems to be evaluated at the start of the run (persumably with the initial value of the Flux F[i]). The initial values seem to always be less than the switchpoint and so only the first if clause is every activated.
Is there something that I'm doing wrong which makes it so that the if-else clause is never evaluated during the actual sampling run?

[fonnesbeck]

Can you prove that the condition ever evaluates to False to trigger the else condition? Also, what is Flux? Data?

[mar-ses]

Yes I believe so, I checked the traces of the F[i]. The Flux = F[i] are just variables that vary as a SkewNormal distribution dependent on i. Most of them should have had the condition evaluate to True.

Also, when I put the print('HERE/THERE') statements in, they only triggered during the first sample and then no more (and all were from True).

I fixed the problem by just making the whole thing a single Stochastic. When I did this, the condition did get evaluated every time and it now works.

def R_logp(value, Flux, A_const, C_const, tauR):
    sw=switchpoint
    if(Flux < sw):
        muR = C_const
    else:
        muR = A_const*log10(Flux) + C_const - A_const*log10(sw)
    return pymc.normal_like(x=value, mu=muR, tau=tauR)

def R_random(Flux, A_const, C_const, tauR):
    return np.random.normal(1.1,0.15)

R = np.empty(N, dtype=object)
for i in range(N):
    R[i] = pymc.Stochastic( logp = R_logp, 
                name = 'R_%i' %i,
                parents = {'Flux':F[i], 'A_const':A, 'C_const':C,
                    'tauR':tauR},
                random = R_random,
                trace = True,
                dtype = float,
                observed = False,
                plot = True, 
                doc = 'Forward modelled radius i')